Applications to Modelling and Predictive Control: Development and Validation of a Semi-physical 1D Model for Virtual Engine Strategy Optimization

点击翻译按钮获取中文摘要

In machine learning, regression is often considered a black box methodology used commonly for identification of suitable functions from a hypothesis. Its primary aim is to estimate a predicted function which leads to minimal expected error on future data, and not necessarily to benefit the overall understanding of the derived output-input relationship governed by the physical interpretation of the system. Consequently, such modelling applications are often highly constrained, allowing training solely for thinly dispersed linear models. To this day, the task of forming intelligent algorithms capable of nonlinear interpretations through physical system observation has received very little consideration, it nevertheless forms the foundation of the field of system identification. Typically, approaches for nonlinear system identification include autoregressively modelling time evolution or utilising Volterra series multidimensional convolution integrals. This project will address the shortcomings associated with the conventional map-based controller design and calibration practices used in powertrain development. It will provide novel, futuristic, non-linear physical causality predictive modelling and experimental approaches for system identification to conclude a real-time capable control system through supervised neural network-based machine learning algorithms for system optimization. The project will be undertaken in collaboration with Koenigsegg Automotive AB and Freevalve AB on the novel cam-less engine technology, Freevalve, enabling major efficiency and power improvements for future powertrains. It will enable the full utilization of Freevalve's potential and the reduction of harmful gaseous and particulate matter emissions, putting the technology in a market-leading position ready for large-scale implementation.